Development of a postprocessing system of daily rainfall forecasts for seasonal crop prediction in Australia

Abstract

There is a need to postprocess seasonal rainfall forecasts from physical climate models to reduce bias, improve skill and restore daily variability for use as input for crop simulation at the farm scale. We develop an extended copula postprocessing (ECPP) method to deal with daily rainfall with numerous zero occurrences. By treating rainfall as a left-censored variable, we derive likelihood estimation and adjust simulation procedure with consideration of zero rainfall occurrences. In a case study for 50 representative agricultural stations in Australia, we test our method to postprocess daily rainfall forecasts with up to 186-day lead time. We demonstrate that the ECPP improves the overall forecast skill from raw rainfall forecasts and outperforms quantile mapping (QM) by checking various verification measures. Though the forecasts for daily amounts are hardly skilful except for the first few days, the forecasts for accumulated totals can be skilful from error averaging and propagating positive skill from short lead times. We also demonstrate that the ECPP can simulate rainfall forecasts with more realistic dry day distribution and daily rainfall intensity than QM. Further research directions including several opportunities to improve ECPP are discussed.